Pavan Turaga
Arizona State University Campus Immersion, AZ·Computer Science
Generative models are machine learning systems that produce new media — images, video, 3D scenes, text — from a prompt, a sketch, or an example. They are improving fast and are being folded into everyday tools, but the dominant recipe is to train enormous models on web-scale data, and on close inspection those models turn out to be unreliable exactly where reliability matters.
- Fields
Computer Science
Engineering
- Topics
generative models
generative adversarial networks
image synthesis
- Grants
- 1 known grant, through Aug 2027
What they work on
- Test whether replacing the fixed input distribution of a generative adversarial network with tunable non-parametric distributions reduces distribution mismatch when generated samples are mixed.
- Develop detail-layering architectures that split image synthesis into a sequence of simpler networks rather than one large model, and evaluate the effect on robustness and model size.
- Build reduced-parameter discriminator networks using orthogonality-type constraints as proxies for physical image variables such as lighting, texture, and deformation.
- Design shape-aware generative architectures that represent object shape explicitly through learnable polynomial basis functions and geometric modeling of the latent space.
More about the research
This project takes a different route for visual media: instead of relying on scale alone, it builds in prior knowledge about how visual appearance works — the physics of light and surfaces, the statistical regularities of images, and the know-how embedded in traditional content-creation practice. The aim is generative models that are smaller, more robust, more interpretable, and modular.
Why it matters
More reliable and interpretable generative models could supply realistic synthetic training images in fields where real data are scarce or expensive to collect, such as manufacturing inspection and healthcare, supporting machine learning systems that can be trusted in those settings. Smaller models built on physical and geometric priors would also lower the computational cost of generating visual media.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Carlos Ramon Ponce
Harvard University, MA·Neuroscience·Early-career
Compare competing ways of predicting how inferotemporal cortex neurons respond to natural images, such as hand-designed features, distance from a neuron's most activating image, and fitted neural networks, in a 'tournament' using chronic microelectrode arrays in macaques.
Yan Yan
University of Illinois Chicago, IL·Computer Science
Develop unsupervised learning models for re-identifying and tracking people and vehicles across a network of city cameras, processing video in real time across edge devices and the cloud.
Alejandro Ribeiro
University of Pennsylvania, PA·Engineering
Develop optimization and statistical foundations for aligning generative models by exploiting that the problem is strongly concave in distribution space even though it is highly nonconvex in the parameters of deep networks and transformers.
The school
Arizona State University Campus Immersion: 36 labs on PI Pages, $1.0B a year on research, 625 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.